GWDG HPC Skill

A Claude Agent Skill for working with the GWDG / NHR-Nord@Göttingen high-performance computing systems — the unified Emmy (CPU) and Grete (GPU) cluster shared by SCC, KISSKI, NHR, and REACT accounts.

It teaches Claude the site-specific details that generic Slurm knowledge gets wrong: which login nodes and filesystems each account type lands on, which partitions you may submit to, how to request GPUs (A100/H100/V100 or MIG slices), where to put data, how to use the Lmod module stack, and how to run Hugging Face models on compute nodes that have no internet.

What it covers

Topic	Reference
Account types → login nodes, filesystems, partitions	SKILL.md (master table)
SSH config, host keys, jumphosts, transfer nodes, rsync / VS Code	references/connecting.md
#SBATCH parameters, job control, QoS, arrays, interactive jobs	references/slurm.md
Full CPU & GPU partition tables, --constraint flags, MIG slices	references/partitions.md
$HOME/$WORK/$PROJECT/$TMPDIR, quotas, workspaces (ws_*)	references/storage.md
Lmod, CUDA hierarchy, conda/miniforge, Apptainer, MPI	references/software.md
Hugging Face caching, offline use, fast downloads, shared weights	references/huggingface_models.md

Install

Option A — as a Claude Code plugin (recommended)

This repo doubles as a plugin marketplace. In Claude Code:

/plugin marketplace add lkaesberg/gwdg-gipplab-skill
/plugin install gwdg-hpc@gwdg-gipplab-skill

Option B — drop the skill in manually

Copy the skill folder into your Claude skills directory:

# Personal skills (all projects)
git clone https://github.com/lkaesberg/gwdg-gipplab-skill.git
cp -r gwdg-gipplab-skill/skills/gwdg-hpc ~/.claude/skills/

# …or per-project
cp -r gwdg-gipplab-skill/skills/gwdg-hpc /path/to/project/.claude/skills/

Option C — upload the bundle

gwdg-hpc.skill is a zipped, ready-to-upload bundle of the skill (the same SKILL.md + references/ under a top-level gwdg-hpc/ folder) for clients that accept a .skill upload directly.

How it triggers

Claude loads the skill automatically when a request mentions the GWDG cluster or its context — e.g. SCC / KISSKI / NHR / HLRN, Emmy or Grete, glogin / glogin-gpu, writing or debugging a Slurm script, choosing a partition, requesting GPUs, managing storage/workspaces, loading modules, or running Hugging Face models on the cluster. SKILL.md stays compact; the references/ files are pulled in only when a task needs that depth.

Repository layout

.
├── .claude-plugin/
│   ├── marketplace.json      # makes the repo installable via /plugin
│   └── plugin.json           # plugin manifest
├── skills/
│   └── gwdg-hpc/
│       ├── SKILL.md          # the skill (frontmatter + main guide)
│       └── references/       # deep-dive docs, loaded on demand
├── gwdg-hpc.skill            # zipped bundle for direct upload
├── LICENSE
└── README.md

Accuracy & maintenance

The partition lists, hardware specs, quotas, and module versions reflect the GWDG documentation as captured and can go stale as the cluster changes. SKILL.md instructs Claude to confirm current values live (sinfo, show-quota, sbalance, module avail, …) and treats https://docs.hpc.gwdg.de/ as the authoritative source. Corrections and PRs welcome.

License

MIT. Not an official GWDG project; the cluster documentation is owned by GWDG.